YM2413 application manual correction and cleanup

So ? Is there a capacitor to change in actual hardware to improve sound quality ?

While normalizing my clipping FM-PAQs (Repro Factory) down to the output volume of my Panasonic FM-PAC it turned out the sound was much more clearer on this modern clone (schematics) where R9 was designed to be 54K but should be 33K instead.
Could any of the people here who I respect as experts (let me edit this one, it sounds like some people earn my respect and some not, this is not the case. I probably should've written respected hardware experts) point me at the wrong capacitor(s) here, if they're wrong?

In my case, please point me at the possible improvement(s) of the konami casing version.

Meits, I am by no means an expert, but the filter in the schematic you linked seems to have a cutoff frequency of ~32.9kHz (R=2.2kΩ, C=2.2nF). That would explain why it sounds clearer than a design based on the erroneous data sheet reference circuit (R=4.7kΩ, C=15nF).

I have a dutch FM-PAC clone, "FM STEREO PAK", does anyone have schematics for that?

This is a huge deal!

Those are Metallized polyester caps. AFAIK they're not that great for audio. Use preferably the polypropylene film capacitors for the best possible filter quality, or if you can't find them just use some new ceramic capacitors instead. Newer ceramics got pretty good these days, and for the SMD designs it might be the only option.

These seem to be metallized polyester caps, but they're bigger because they're for 100V. It won't fit on some smaller designs.

Does this effect existing software implementations (like e.g. the one in openMSX)?

@sd_snatcher You already fixed/modded some hardware? Would it be worthwhile to demonstrate the difference?

About sound quality, OCMs do not have mixer and many slot expanders too.

This might be a silly question, but how about the modern FM-pac replacements, e.g. Fmpac Lite?

boomlinde answered that a few replies up

i dont like the crisp sound . i leave my fm like it is

Yes, though only indirectly because the resampler needs to have it.
But there are (currently) no specific per sound chip filters emulated (nor any other analog effects).

Sometimes it's just unforgivably bad that what you hear has nothing to do with what you should hear: Philips, Sanyo, etc.

FmPac Lite and every other FmPac clone have the real thing. The YM2413 chips are cheap and easy to find, much cheaper than FPGA. Also more authentic sound. Only systems and extensions which are already FPGA-based for other purposes like 1chipMSX and GR8NET have an FPGA-based YM2413.

@Lord_Zett

Don't worry. Fixing the filter doesn't make it loose the 80's feel just like the SCC has a sharp sound and still sound like 80's sound synthesis. The quality improvement is more like switching from the RF video connection to an RGB video connection: It's still 2D, it's still the very same MSX, but now you'll be able to enjoy all the little details that were missing.

FRS mentioned to me earlier that most clones also have a LPF cutoff value which is too low.

Did someone apply this yet? Not too many Sanyo Wavy 70FD’s out there I guess, but I’d love to hear a before / after comparison of the difference. And I hope a similar fix will be published for the turboR GT, because I’m super curious about what it will do for the Illusion City intro .

And that one can also be interesting to compare to:

1. openMSX
   

If you record with opens on 44khz, which resampler did you use?
You can also record all channels separately at their native frequencies.

Perhaps it's not that useful to show all these inaudible frequencies?

I'm no signal expert, but maybe there's something to understand about the underlying sound processing here with those high-frequencies (maybe aliasing stuff ?). I prefered to leave everything intact and un-processed, and will just wait for some experts analysis, I would be happy to learn new things again here !

Was going to post that, but just found it: set frequency :) ! Will redo my openMSX recording...

EDIT: I just re-recorded everything using "record_channels all" (thanks Ren), then combined all different WAVs into a single 192KHz FLAC file with Audacity (mixing is done at export time by Audacity, not sure which method it uses for that purpose). Frequency Analysis diagram and FLAC file for openMSX have been updated above, so just ignore my comment about 44KHz.

Your openMSX recording is totally distorted now. I believe (read my experience as well I described in my post above) you can't simply mix the channels in some app and get the same result as openMSX's mixing/rendering.

-edit:

Recording openMSX with "soundlog" should be fine. That's also the output that openMSX users will normally hear. Just make sure to select the hq resampler (high-quality) if you care about sound quality. The output frequency (44.1kHz vs 48kHz) doesn't really matter as long as you only look at the audible part of the spectrum (below 20kHz).

Also for the other graphs I'd recommend to only look at the part below 20kHz. Even if you sampled those at 192kHz (I'm assuming you're using a PC sound card to sample, not some expensive high-end audio component). Because many sound cards assume everything above 20kHz is anyway inaudible and make use of that freedom. So the stuff above 20kHz likely shows (more) characteristics of your sound card than of the thing you're measuring.

It would be good to include a graph of a 'silent' FM-PAQ output (or other hardware). To get an idea of the noise floor of your setup. Everything below that should also be ignored in the comparison. (But this is less needed if you cut the graphs at 20kHz).

Hi Meits,

No hardware changes at all on the tests published here (also no slot expanders used). Thanks for the tip on the R9 resistor, the volume is a bit louder on my FM-PAQ lite (you can hear that in the FLAC files here), but I noticed no clipping yet.

Did you disable the internal turboR FM-PAC on your ST ? (or maybe you're just using the FM-PAQ output with no PSG?).

AY-3-8910 and YM2149 have 3 outputs (1/channel), the S1985 MSX-Engine’s PSG has 2 outputs (left/right), the S3527 MSX-Engine PSG has a single mono output. OPLL has 2 outputs (FM / Drums), and the internal digital mixing does not sum, but uses an unusual (and cost-saving) time-division method.

All of these are mixed together. If you’d plainly 50/50 mix the PSG and OPLL channels, you would get very different mix balances even just depending on the PSG type. So it’s not just a matter of summing all channels, there’s different mixing circuitry in different places.

OpenMSX has per-chip mix settings in the machine and extension configurations iirc. I submitted a patch once to bring the SFG-05’s YM2151 volume more in balance with the rest comparing by ear on the real machine (tricky and a bit unscientific, but it’s way better now than it was before). That chip has a massive dynamic range btw, very easy to make it clip, so it has to be mixed carefully.

I use its jack output and the custom jack output on my mfrscc+sd2 with a cloned PSG (when desired).
The FM-PAQ lite is too loud compared to any other soundchip I have. Every mixer I connected it to had the clipping led flashing. Two hardware gurus around here (Daemos and Omega) calculated the exact same desired value of 33K in stead of 56K. It sounds smashingly good and the clipping leds on my mixers stay off now.
If you're opening the device, please check the capacitors. I found one which was too weak on voltage according the schematics and one was placed the wrong way around. Might be just mine, but it stays handwork

Thanks for your input Grauw (@Louthrax: seen it as well?)

About the mixing: it is somewhere here: https://github.com/andete/ym2413/tree/master/results I hope Wouter wil pinpoint to you where it is described.

Wow great post thanks you very much!

as someone who doesnt know the original tune (yeah I know I should be ashamed but thats another topic) it is great to hear the differences.

While i am not professional sound engineer but just an occasional wannabe music maker, i noticed this:

- Original unmodifed FM pac wins. Not only because the song was probably made on that device at the time, but sounds the most balanced and most clear to me.
- The sony: I have seen in various topics ppl saying it is "great or perfect sounding", but I feel it sounds way too bassy, I dont feel it is correct.
- FM PAQ lite has weird distorted sound
- Turbo R ST sounds thin. I guess this is the lowpass filter thing mentioned a few years ago in some topic
- Openmsx sounds so different :/ Kinda sucks since I primarily use openmsx when i make music. Guess I need to get some better speakers for my real msx setup and use that for making music instead.

However it is sorta hard to judge properly because I only listened to this for about 30 minutes and this is psg + fm pac mixed.

I read some remarks here that the FM-pac needs fixing? tbh I don't think it is needed since imho it is the best sounding one of the bunch. However I dont know how it sounds on other ppl's setup? maybe there are different revisions of the fm pac?

The openMSX sound (without recording issues) is somewhat close to the sound of Panasonic MSX2+ machines, except for the low-pass filter on the FM sound, of course.

I'm sorry sd_snatcher, but you are WRONG!

Let me explain.
The yamaha doc speaks of 20kHz cutoff, but this is for a BRICKWALL filter.
This is a filter with a very steep roll off.

In fact, since the stated samplerate is 50kHz, if this filter is applied at 20kHz it should filter out just about all sound above 25kHz. It would only have 5kHz range (between 20kHz and 25kHz) to do this. Very sharp.

BUT. The filter in the schematic is a simple RC filter. These filters only have a slope of 6dB per octave.
If you applied such a filter at 20kHz then it would not filter out enough step noise and aliasing.
And THIS is the reason the cutoff of this simple filter was brought down to 2kHz.
At 2kHz the filter manages to suppress enough garbage above 20kHz and so does its intended work.

Now, IF you want to correct this (but i don't honestly see why you would do this in the original fm pac because it sounds beautifully warm as it is and has a character) then you should design a proper brickwall filter with, i dunno, 90dB/octave or more. Remember, at 20kH you only have 5kHz to filter the sound due to niquist rate.

I’m looking forward to the discussion . I hope it can be done with cool heads (maybe not use unnecessary hyperboles), since I would really like this topic to be discussed properly.

It’s a bit hard to participate myself since I haven’t performed the mod, if I had I would grab the oscilloscope and compare with figure III-b of the application manual…

"It’s a bit hard to participate myself since I haven’t performed the mod, if I had I would grab the oscilloscope and compare with figure III-b of the application manual…"

That figure got me thinking...
Because two channels are muxed (and something about drums being doubled) and the sampling freq is 50kHz, it could be possible that the optimal cutoff of an ideal brickwall is an octave lower than what i suggested. That would mean a cutoff of around 10kHz instead of 20kHz.
Unfortunately i don't have an fmpac here to test on.. Maybe the muziek module has some answers

So this is my current understanding;

The OPLL does not have an external DAC and instead outputs the values for each channel in sequence at a high frequency of 447 kHz, with each cycle completing at a rate of 50 kHz, and it relies on something external to smooth this out (effectively mixing).

This 447 kHz output can actually be connected directly to speakers (as mentioned in the application manual), since no speaker can reproduce this kind of frequency they will naturally smooth it out, however it will produce some artifacts due to the speakers being driven at frequencies for which they were not built (this is part of the 192 kHz music debate), so for this reason the result will sound more pleasant when there’s a low-pass filter inbetween which dampens the frequencies above the audible range.

So the biggest question is, what amount of filtering is needed to sufficiently smooth out the OPLL output so that it ends up as much as possible as the sum of the channels, and the high frequencies which speakers do not like are subdued, without affecting the audible frequencies (the actual music) too much.

Analog filtering I don’t know much about, but for digital filters I know it’s actually pretty hard to get an “ideal” filter; if you cut off sharply, it causes peaks in other parts.

The Music Module uses a DAC IC unfortunately, like nearly all the other OP* chips, so it’s not really comparable I think.

"So if I understand your view correctly, the current proposed fix leaves the filter too "open" (not supressing higher fequencies enough)?"

Correct.

"The OPLL does not have an external DAC and instead outputs the values for each channel in sequence at a high frequency of 447 kHz, with each cycle completing at a rate of 50 kHz, "

Aah, ok, that clears up the muxing.
So in case of a final 50kHz samplingrate you can have a filter with a cutoff of about 20kHz.
Artefacts start at 25kHz.

"Ideally, with an ideal brick-wall filter (which does not exist), only the inaudible frequencies would be cut off, and all the audible frequencies are passed through."

Yes, no, sometimes.,

The real objective is to filter out everything above half the samplingrate. This is a requirement stated by the sampling theorem.

But since this hardware is designed for humans to listen to the samplingrate will be around twice the highest frequency a human can hear. In this case the chip outputs a 50kHz sampling stream.
50kHz samplingrate can carry a signal up to 25kHz. How convenient!
And since we can't hear above 20kHz you get about 5kHz of free space to make the filter work its magic.
Strictly speaking the frequencies between 20kHz and 25kHz are still in band for the DAC. But since we can't hear them they don't matter.
Frequencies above 25kHz are not the intended output of the DAC, so they should be made unhearable.

So the filter should start filtering at about 20kHz and be sufficiently done by 25kHz.

By comparison let's consider a CD.
CD format claims frequency response up to 20kHz (human hearing limit).
It has a samplerate of 44.1kHz.
So the highest frequency that can be represented is about 22.05kHz.
Dynamic range is stated as 96dB.

So in CD players there is a brickwall filter that takes the signal down by 96db in the span between about 20kHz and 22.05kHz. This is very very steep!

To some degree you're right.
But i always thought the rounded top end of the FMPAC was a nice feature.
30 years ago I used to have a MM instead of the FMPAC and i always was a little bit jealous of FMPAC owners because it sounded a bit sweeter.
No doubt this filtering was the cause.

This cannot be understated. This is especially valid for 'bedroom' composers (of any kind) with non-reference audio equipment and acoustics.
Tho if anything, a typical analog path (amp, speakers) would probably have had even more high frequency roll off then what is suggested in the doc we are discussing.
We are talking late 80's early 90's here and i bet most sets in those days couldn't reproduce 20kHz at hearable levels, especially the ones we had in our bedrooms.

In that case i would advice to use an equalizer (preferably a corrective eq with 32 or so fixed frequencies) to set the sound to ones liking. This will not screw with the actual reconstruction of the samples but simply lift the hearable range that has been subdued by the cheap RC filter design.
I think you can pick them up for pennies these days.

That is actually too specific. The filter in the first place is there to make sampling work.
If the only objective was to mix the muxed channels then a filter in the 223.5kHz range (half of this 447kHz) would also suffice.
But such a filter would not suppress any of the out of bound frequencies generated by the 50kHz sampling rate of the combined channels.
So assuming the lowest samplerate (which turns out to be 50kHz) you need to filter below half this rate, so all filtering needs to be done below 25kHz.
And once you satisfy this criterium then the criterium for the filter for the 447kHz frequency will automatically also be satisfied.

This muxing business is a funny byproduct of sampling theorem and that is why it 'just works' IF you follow the stated 50kHz sampling frequency.

So all you need to do is concentrate on the 50kHz sampling rate and the muxing will work itself out.

@redman

I'm well aware of what step noises are and what's the objective of the reconstruction filter. I'm also aware that any solution must take many factors in consideration (like the cost/benefit and how complicate the surgery would have been).

You seem to be well versed in filters. In that case, would you kindly calculate and plot here the Harmonic distortion curves in dB for the three filters, so others can see how much distortion you're taking about?

1) Old low-pass filter from the YM2413 datasheet with a 2.26KHz cut-off frequency

2) New low-pass filter from the YM2413 datasheet with a 22.6KHz cut-off frequency

3) New 20KHz low-pass filter you propose (with the corresponding circuit of course)

Doing the calculations, the original filter would be at -22dB for 25kHz, and the new filter would be at about -3dB for 25kHz which is definitely not good. Using a third-order butterworth filter, you could make a filter cutting at 6.8kHz for the same -22dB at 25kHz from the original filter.

@ricbit

1) Are you saying that the THD for a filter that causes suppression of frequencies in the entire very audible range from 6.8KHz to 14KHz will be smaller than the THD of one that leaves up to 2% of THD from aliasing in the almost non hearing frequencies >16KHz?

2) Nobody is connecting their headphones directly to the MSX audio lineout (there's no power for that, BTW). This means that you have to connect it to a decent amplifier. As a consequence:

2a) Any decent modern sound equipment has its own 20~20KHz filters that will be way more sophisticated than anything we can build with analog circuitry inside the MSX. This means the best approach is to just "get out of the way" and distort the sound the least you can, because otherwise the sound equipment won't be able to recover the suppressed frequencies. But it will do the anti-aliasing wonderfully.

2b) If you connect your MSX to a not so decent sound equipment like some "multimedia speakers" or the TV sound output, those will muffle the sound naturally, as they just can't reach the higher frequencies that would show aliasing anyway.

@Grauw

In the audio output analog domain, aliasing presents itself mostly as high frequency "noise", where "noise" in that context is anything that didn't belong to the original wave. I.e.: Where a perfect sine should have been, there's now a ladder of low resolution that resembles the original wave, but its sharp square borders result in high frequencies being emitted.

Many times, this noise is outside the human hearing range, or at least sitting in the barely audible range of 16KHz to 20KHz. If it gets too strong in frequencies lower than 16KHz, then your sampling rate must be really terrible and it will sound like almost constant sharp beep in the background.

This is why the reconstruction filter is also called "anti-aliasing filter".

In that case, if we’re going to rely on external equipment, why have a low pass filter at all .

I think the a low pass filter is present to make the output consistent instead of relying on the capabilities of the audio hardware connected to the MSX?

Question is, how much does the low pass filter need to filter, and how much can we rely on the characteristics of external equipment. Sure, a brick-wall filter would be ideal, but is the 20 kHz RC filter not a better compromise than a 2 kHz RC filter which greatly affects the audible range? I certainly do not like the latter side effect, would prefer to minimise it.

p.s. I’m not familiar with analog filter design; Is the cut-off frequency a direct result of the capacitor value, e.g. using 0.003 µF caps would produce a cut-off frequency of 10 kHz, or do resistor values also need to be changed to match?

Ok, great, I understand. That’s handy to know in case I want to do some experimentation of my own . So I could replace all components except C10 and C17 for the improved gain / SNR / mixing, and then experiment with a few different capacitor values to adjust the RC filter, right?

Did you experiment with different capacitor values? Maybe a 0.003 µF value (10 kHz filter) has barely any negative effect on the audio, while improving the reconstruction filtering?

Since it is easy to notice the effects of the filter on the audible domain, while it is difficult to judge the effect of possible imaging/aliasing/whatever (since it really depends on the your audio system), I could imagine it’s best to try to try to put the cutoff filter as low as possible as long as it has no significant effect. There are many values inbetween 2 kHz and 20 kHz…

The more the marrier.
But analog brickwall filters are problematic.
I think that for something like this chip you will have to live with some compromise one way or another.
But i also think that a lot can be improved over a simple rc filter.

Not from a sampling point of view i think.

Yes, aliasing is the foldback of images into the audible spectrum when sampled.
What is left after a non filtered DAC output is actually these high frequency images.
If you look at a spectrogram from low to high frequencies what you actually get is first your normal spectrum, then a reverse frequency copy of the spectrum, then a non reversed copy, then a reversed copy, etc etc,

If you would sample this signal without an anti aliasing filter then you would indeed get aliasing.
But actually any non-linear process in the path would reflect these images back into the hearable spectrum. If, for instance, an op-amp is a bit slow then it will create aliasing from this signal.
And most audio op amp circuits are not optimized to handle ultrasonic signals.
You don't want that stuff to hang around on top of your signal. It can (especially with classic hardware) interfere with all kinds of things in the audio path which leads to worse sound.

How about this filter IC.
It's an 8th order filter that goes up to 25kHz. Costs about $5 in bulk and runs off of 5V.

And i'm sure there will be other commercial products out there that could be used if better-then-RC filtering is wanted.

You are confusing concepts. There is aliasing, there is quantization noise, and they are different.

Aliasing happens when you sample an analog signal with a frequency less than twice the maximum frequency in the original noise. If you have a sin of 5Khz and sample it at 2kHz, you will hear a 1kHz tune because the frequency at 5kHz gets aliased to 1kHz. You use low-pass filters before the sampling to avoid that.

The filter above is not a reconstruction filter. A reconstruction filter is needed when you convert from the digital domain back to analog. Reconstruction filters are hard to design and they depend on how the reconstruction is being done. I bet OPLL uses a zero-order hold because it's easier to implement.

After a zero-order hold, a perfect low-pass filter filtering at the nyquist frequency can reproduce the signal exactly, but unfortunately a perfect low-pass filter doesn't exist. Imperfect low-pass filters will give you aliasing artifacts which have nothing to do with noise, instead you'll have copies of your signal replicated above the nyquist frequency. This is the anti-alias filter, you're removing the extra copies due to aliasing.

Low resolution is another matter entirely. PSG has 4-bit volumes, CD-ROM has 16-bit volumes, and no finite amount of bits will be enough to reproduce an analog signal. When you sample the signal with a finite amount of bits, the end result is quantization noise. This is not the same as aliasing, because aliasing you can remove with ideal filters, but there's no ideal filter in the world that can remove quantization noise (however there are very recent advances in the field where you can train a neural network and use deep learning to guess the remaining bits - this is well beyond the scope of OPLL filtering however )

There are some very nice courses in Coursera where you can learn this math, look for "digital signal processing" courses there.

Now back to MSX. If you are going to rely on the fact that modern audio circuits have sharp filters at audible range, then your best choice, as I've said before, is to remove entirely the RC filter. Short-circuit it. Why bother adding a filter that will only attenuate your band-pass; the high frequencies will be filtered externally anyway.

For the purpose of comparing "integrated circuits", or "reconstruction filters" with DAC vocabulary:
1.
C12 and C13 at 0.015uF and R7 and R8 at 4.7kOhm in Panamusement FMPAC SW-M004.
2.
C10 and C11 at 0.015uF and R29 and R30 at 4.7kOhm in Panasonic FS-A1WX.
3.
C10 and C17 at 0.033uF and R13 and R17 at 4.7kOhm in Panasonic FS-A1QT.

Nice one!
So those are basically the same on the fm side filtering.
It is also noteworthy that there is an influence from the internal amps and mixer in a given msx. There is added distortion and there can be more filtering happening there as well. Yummy. :)

No, they are the exact same values as in the app manual and they resolve to a cutoff of 2.26kHz .
It's just that the mo and ro outputs are filtered separately so you get two RC filters for the chip.
Interesting option is to filter rhythm differently from melody.

Thanks for confirming, I thought my ears were the problem. Glad to hear (pun intended) that I'm not the only one. So imo this is something that could be looked at.

compared to what? Other computers with an ay-3-8910?

I'm sorry, but the one who's wrong here is you. You're reading more in the YM2413 datasheet that is there. It just requires a low-pass filter to mix the multiplexed ouput of the chip and nothing more than that.

The YM2413 is a low-cost chip. A BRICKWALL filter would have been ridiculously expensive and out of question for this design. A sane engineer of the era would laugh at that proposal and tell you to either live with some distortion/imprecision or get a more expensive OPL chip instead.

I haven't been answering your arguments just because they're incomplete, inconsistent and biased towards your personal preference for muffled sounds. And you clearly express that though "Yummy" and "the more the merrier" expressions along your argumentation. You haven't even shown the distortion curves, as it would have shown that the distortion you're talking about would happen outside of the frequency response of the entire circuit.

As I already explained to ricbit, the argument about nyquist artifacts for a 50KHz sampling rate inside an 80's circuitry is just out of reality. Those high frequency artifacts ("noise", or "duplicates" if you wish) won't ever survive the natural massacre of high frequency killing components that follow up in the circuitry.

This means your low-pass filter for analog reconstruction is already present and you just don't need to add another one to suppress even more the high frequencies, unless otherwise that's your primary intention and not the reconstruction.

That's 80's design being presented to you. You take advantage of any analog characteristic that's already present on a component or circuit to save costs. The whole YJK color system (also designed by Yamaha) was also based in that principle: the chroma "reconstruction filter" is done externally to the VDP simply by relying on the fact that the analog circuitry of NTSC has a low-pass filter behavior.

The whole composite artifact colors is based on the same principle and was widely used in 80's computers/video-games. If anyone proposes to add a high-precision low-pass filter to an 80's design to be able to enjoy this trick, they clearly didn't understand how the whole thing was envisioned to be cheap since its inception. And yes! It's based on sending untouched high frequency signals to a circuit that has a low frequency response so it naturally acts as a reconstruction filter for you.

Your arguments are so inconsistent that one time you propose a redundant brickwall filter in order to have "the perfect sound reconstruction", while OTOH you vouch for an absurdly strong artificial low-pass filter just to suit your tastes for muffled sound.

This is like forcing everyone to apply strong blurring filters on their LCDs, just because you don't like the fact that they show pixeling. When just a soft scanlines emulation would be enough to achieve the necessary effect without any of the blurring.

So, if you want to point fingers and tell someone that they're "WRONG, WRONG, WRONG" is this topic, point your finger to yourself. You seem to have a good knowledge about theoretical signal processing, but your approaches and assumptions for 80's analog circuitry are totally incorrect.

Even in the 80's you could make better filters than the one pole RC filter.
CD players were there in the 80's and they had better filters than RC...
But i get your point. For a chip with such a grated high frequency spectrum you don't want to spend a lot on the filter alone.

You are assuming a great deal of things right here.
I heard the examples you recorded.
And that is apparently a situation where the signal went through all the 'killer' components that you mention here, i hope you agree.

Well, despite all the filtering you assume is there, it sounds like absolute crap!
Seriously, way way WAY too much high frequencies. And it is terribly noisy. Not normal analog noise, mind you. All kinds of spurious crap on top of a haze of a modulated high freq.

Whatever you did to the audio path on that recording it didn't improve anything.

In fact, i'm not sure you actually hear the whole spectrum as these piercing high frequencies are not pleasant to listen to at all. They completely ruin the fm pac sound for me.

And apparently you recorded this off of a real msx.
So all these arguments here about the audio route and components in the msx are nonexistant.
What you recorded is the actual, filtered by bad components and tolerances, and it still sounds like a mess with way too much high frequency content.

So that is not really an argument of any kind.

(Also, coupling capacitors act as HIGH pass filters, but that is besides the point )

These 'passive mixers' with an op amp behind them are actually pretty decent mixer designs.
You'd have to work out the math, but it does work.
In fact, these kind of mixers are a common circuit building block and they are not passive.
Without the op amp they would not work correctly. So these are actually Active mixers.

Also, transistors have a linear range that is suitable for amplifiers.
When it's not linear you get harmonic distortion which doesn't sound bad to the ear in small quantities.
This is not a problem in and of itself.

You shouldn't concentrate of the theoretical side of this.
You should actually listen carefully to the results you produce by changing these cap values.
Then you would hear that your 'solution' actually degrades the sound of the OPLL as it comes out of the MSX.

That is simply untrue as you yourself demonstrated with your recording.
The modded OPLL sounds like garbage (sorry to put it so sharply) and it needs more filtering to sound decent.
It is very obvious.
Well, it is obvious to me, i don't know about your ears :)
And again, i base this on your recording.
You yourself have given the proof that the caps should stay in the range that was originally specced by the application notes.

And we're not in Kansas anymore, Dorthy. ;)
No seriously, we have better and cheaper stuff now.
Tho i'm starting to doubt that any kind of better filtering won't help the case much.
I looks more and more like there is severe high frequency crap being generated in the hearable range. No ammount of filterting would fix this.

You have to consider this:
Why did all these super professional mutlinationals all decided that 2.2kHz filtering was the best option (except the people from PAQ)?
Are you assuming that they were lousy engineers that worked blindly without testing anything, without hearing the result?

I mean, you really think that these people would not have considered raising the cutoff of the RC filter?
Did you ever consider that they had very valid reasons for not doing this?

Did you ever consider that -gulp- YOU are the one who is wrong in this case?

Of course you didn't.
You actually had the nerves to revise a design document that served the whole industry for years.
You must assume that these people are stupid fools that didn't manage to correct their mistake for the whole period of this chip being on the market.
I would say that this is pretty arrogant.

I mean, i wouldn't have minded if this was just a simple mod to make the OPLL brighter then designed.
But the fact that you pose it as a solution to an error really ticks me off.

Yeah, yeah, you can make up analogies all day long. :)
Fact is, the cap mod sounds terrible, even despite all your claimed internal filtering. It is not the right way of doing things. Unless you're deaf.

Yeah, yeah :) Ok, a little lighter.
But this analogy doesn't hold.
If going from 2.2kHz to 6.7kHz gives you more scratches then surely putting the Fc at 20kHz exposes many more scratches.
And this seems true.
By your own audio example the sound seems unbearably bad.
I honestly can't see how you would call that an improvement.
Never mind calling the original design of the filter a mistake.

You are actually kindof demanding that your opinion (which is arbitrary) on what is best is be made into truth.
I bet you didn't even contact the writers of that document to verify your findings.
You just put it there as if you are the sole source of all truths.

IMO you should not have touched that document.

Now look, to understand what can be done you first need to understand what needs to be done according to theory.
So you get to play with ideal situations and ideal components.
Then you apply this ideal to reality and see what survives.
So its only natural to first look at the ideal situation.

In any case, i just want to stress that your revision to the application manual is an arbitrary opinion of what should be done. It is not a fix, you actually change something that was put there by design.
If you do not follow the original values then you are left with a horrible sounding chip.

Yamaha, Panasonic, Sony and others that have implemented this chip must have realized these were the correct values because otherwise they would have changed it themselfs!

If you think you're a better and more professional audio circuit designer than the engineers at these firms then i think you're full of air.
I mean, come on, Yamaha doesn't know how to choose cap values? You've got to be kidding me, right. :)

And, given the result, i wholeheartedly agree with them. They put the right values in the application sheet, and for good reasons.

This is because i changed my opinion about the chip.
At first i though it would behave nicely if filtered properly for its sampling rate.
Then i heard the actual result of the 20kHz mod and i changed my opinion. The chip itself is so noisy within the hearable range that i consider the whole 10kHz and higher region as severely compromised. This problem can apparently not be fixed at all. So the only sensible thing is to filter it out.
The low cutoff of the original design seems to do exactly this.

LOL, at least i'm not the one revising design documents based on my opinion. Think about that one for a minute.

I'm not assuming anything. I have looked into the circuits personally and created the recipes to modify it, remember?

You're the one that's assuming that there's absolutely no other filtering in the audio path.

They sound exactly as the other OPL chips with more elaborate DACs, like the OPL3 and OPL4.

So it's just your personal taste that's the issue here.

You're only right about the part that I recorded it on a real FS-A1ST. But you're again making false arguments with nonexisting assumptions based on circuits you never bothered to look.

As I said before, the FS-A1ST has a 20KHz lowpass filter integrated to its global output preamp.

Even when they have coupling electrolytic capacitors associated?

So let me tell you that there are *a lot* of MSX audio circuits that don't have the opamp. Go check the NMS-8250 service manual.

A Double standards fallacy in front of everyone? A single transistor configured in a way that produces saturation of a lot of frequencies has acceptable distortion, but a lighter filter introduces an unbearable distortion.

You are entitled to your opinion. That's your personal taste alone.

Now it's the argument from authority fallacy.

Since your arguments and assumptions have been proved wrong, you're now resorting to a lot of fallacies.

And now the Argument from incredulity fallacy.

You're assuming the opposite. If the Turbo-R crappy sound just doesn't prove that for you, nothing will.

You're repeating this since the beginning, but you have yet to prove it. Just the ad-nauseum fallacy won't hold your ground here.

If you want to prove you point, go ahead and show some solid proof. Otherwise it's just your personal taste and opinions about the sound recordings that are being exposed here.

I'm under the heavy impression that some sacred ground was touched.

Are you serious? The whole lot of 70's and 80's consumer electronics fall out of this law you're creating now.

Another argument from authority arguments. You idealize to much and forget about how these projects were rushed and cash constrained. Most of the circuits were just copy/paste designs from the datasheet, without too much time being spent on it or not even small optimizations.

If you had analyzed more circuits, you would have known that Sega used a completely different mixer for the YM2413 with fewer components and different values. Were Sega's engineers wrong and they should have blindly copied the datasheet circuit as the perfect men from Yamaha, Panasonic and Sony?

If you had bothered to analyze even more circuits, you would have known that Sanyo tweaked the value of the low-pass filter because they obviously noticed that something was wrong. And that many MSX music cartridges have also done that. Are you saying that all those engineers are stupid?

And now, the only fallacy that was missing. The ad-hominem.

Given the huge amount of fallacies being used here, I consider you have lost your reasoning and this discussion has ended. Nothing productive will come out of this. I have more productive things to do with my scarce time. Have a pleasant day.

Ok, as the author of the MSX Assembly Page I’m now going to say what I want and will do.

I’ve given the original datasheet back a little more prominent place, alongside FRS’s filter improvement rather than as a bit of a sidenote. Hopefully that addresses some of the concerns.

From the actual useful discussion here inbetween all the crap I now feel that is the right thing to do, since I’m not convinced that there is a decimal point error as FRS suspects. There is theory and there is fact, and the MAP is about facts, so I do not want to declare it as canonical unless it’s proven. This is an editorial choice of me, the MAP author, who strives for accuracy and completeness of information above all.

FRS, I hope you understand my position in this. I wish I could do this when people felt less polarised. I linked to your fixes page so you can elaborate there as much as you want, also if you have suggestions for improved presentation on the MAP or a v2 of the AM please let me know.

I do think that reducing the low-pass filter makes the OPLL sound better though, subjective as that may be, and will try this out myself and of course report back when I do, because there’s nothing like hearing it for yourself (Soundcloud’s compression is pretty crap). I say that as someone appreciating the wonderful crisp sound of the YM2151 and YM2608. Thankfully thanks to explanations in this thread I now know that I can easily tweak the filter cut-off frequency, and I feel a bit safer modding my turboR.

As I said before, I think it would be good if some of these explanations and nuances made it to the description of the fix and the YM2413 AM, because I feel it would benefit the MSX users who want to fix their turboR’s audio and remove some of the controversy by giving them a choice. But that’s up to FRS, who is the only one who is contributing something to the MSX community here.

Also, sorry for suggesting going to Coursera, since you asked us to draw the response of the first order filter I just assumed you didn't know how to do it. My bad.

I agree with this in theory.
I still feel the chip requires its own filtering but if one really wants to depend on the playback chain to filter then no filter is the best option.

Yes! This would be very interesting.

It wasn’t even clear that there would be debate until after it was published. I certainly did not expect that anybody would take so much offense to it.

FRS is the author of that particular cleaned up version of that document. I can ask for changes, change the context in which the information is presented (which I did, better now?), or choose to not publish it.

That’s what this discussion is for, right. Parts have been very informative. For me it is also important to understand the subject matter better, both because I want to apply the fix and because I’m the curator for the quality of the information on the MAP. So I was hoping others would pitch in. As part of the outcome, things can be updated or clarified.

Why would I be picking sides by publishing a document? I don’t even want to think of sides. There were no sides to begin with. I am in a split position because people felt the need to turn the discussion sour, there was ABSOLUTELY no need for that. What PAC said.

Sure, maybe this is not the best format. The same thought crossed my mind, but truthfully, it’s fairly irrelevant. I don’t understand why you put so much weight on it. It’s like, whatever, it can always be changed if it is contentious, but not really important at this stage.

The topic of the actual filter change itself is many many times more interesting. Now the strong and weak points have been identified, as ricbit nicely summarised. I would love to see some further tests performed, some people implementing it and sharing their results, and to look at the results with an oscilloscope. Maybe you can do it? Turn theory into practice?

The high harmonics are affected quite a bit, because the attenuation starts to accumulate at such a low frequency. I think already doubling the cut-off frequency will have a very audible effect.

Yes. Here the concerns were twofold:

Some people expressed a dislike of the lack of attenuation of audible frequencies. What puzzles me there is that neither MSX-AUDIO, OPL4, OPM or even the OpenMSX OPLL output have such a low cutoff filter, none of which I’ve heard many complaints about (certainly not in terms of “ear bleeding” and “garbage”).

Secondly there is the question whether residual inaudible frequencies will negatively affect the audible signal due to interaction with components between the OPLL and the speakers. FRS’s experiences say no, but there are doubts, and it would be good to get some more experiences and confirmations.

At least for me it currently has the status of “theory” rather than “fact”.

The problem with that question is that it is both sensible and extremely difficult to answer, not only because who’s going to interview all those developers, but also because there are many more factors which affect the sound than the filter cut-off alone, such as the machine circuitry and the audio equipment used, even the room that he was in, curtains or not, etc.

However the simplistic answer is, 2.26 kHz most likely.

About the turboR, it has a cut-off of 1.03 kHz, so definitely way too low, and the mixing is off completely as well. That part of FRS’s audio fix has no contention at all I think. To be clear, that’s FRS’s fix with 0.015µF caps.

Increasing the cut-off above 2.26 kHz up to 22.6 kHz (0.0015µF) or something inbetween seems to be up to personal preference, although that preference has so far only been expressed based on FRS’s examples, which do not compare purely the OPLL filter change, plus they are in crappy Soundcloud quality. Once people implement the fix, maybe experiment a bit with cap values, hopefully we will start to see some opinions which are formed on a better basis.

However, as I mentioned before, if you don’t hate the openMSX OPLL sound, or the OPM, then I see no reason why you would dislike the output of FRS’s improved filter. Assuming there is no distortion introduced due to residual inaudible harmonics.

So there are a few approaches.
One is to leave everything as it is, with a simple RC filter at the designated cutoff.
Another is to increase the cutoff.
Yet another is to improve the slope of the filter and then increase the cutoff.
I suggested an 8th degree low pass filter IC that would do about 48dB per octave (compared to 6dB per octave for the RC filter)

The approaches have different outcomes.
Approach one leaves everything as is so the outcome is known.

Approach two will depend on the chosen cutoff but to my ears a too high cutoff introduces noise that is in my opinion unwanted. In fact, it makes the sound unpleasant for me.

Option three kindof an unknown. A lot depends on the actual quality of the digital signal that is generated inside the chip and the general quality of the DAC.
If the high frequency noise is only the result of ultrasonic information (which is generated by the DAC as it has a step response) then a steep brickwall like filter would help to clean up the noise and leave the intended signal alone.

If, however, there is too much internal noise that is already projected into the hearable range (up to 20kHz) beforethe signal leaves the chip then there will be no way to improve the sound much by using a very steep filter.
A filter that is a bit steeper than a RC filter could maybe still allow you to better control the amount of high frequencies you're willing to put up with but the rule then becomes the more you let the high frequencies pass through the more noise you will get for free. The noise would gradually get introduced in the top octaves of the hearable range, so about 5 to 20kHz.
If this is the case then a brickwall filter would be mostly useless as the noise is already built in and not the result of improper reconstruction of the samples.

I base this on what i hear in audio files of opll with a high cutoff RC filter but i don't actually have the chip so i can't properly test this. So it's only a theory.

The upshot is tho that if this is true then yamaha engineers would have had a very good reason to filter the top end as severely as they did.

Nice! :)

Could you maybe post the specs of this recording?
What are the frequency of the tone, samplerate, bitdepth?

The spurrious noise peaks could have various reasons, from chip artefacts to noise on your recording chain. They seem to be smaller then the dynamic range (horizondtal step size) of this signal so let's not worry too much about those for now.

What is maybe more interesting to analyze first is wether the horizontal steps we see here are caused by the DAC or by the waveform in ram (and the actual dac resolution is higher).

I also want to note that a 1/512th change to the amplitude is still pretty significant. This comes down to about 9 bits of resolution which means that the noise floor will be about -54dB below full scale. That's definitely hearable as noise but still pretty decent.

Anyway, time for bed., :)

All the details should be here. It was recorded in a controlled environment with a lower than usual clock rate. The stepping is due to the waveform. Since YM2413 OPLL outputs every channel individually, the DAC resolution directly reflects the internal representation.

But maybe it’s easier to just check the YM2413 datasheet :D. It says the noise level is -63 dB (page 26).

The turboR ST has that filter on the op amp as well, see page 21, right after the mixing. So, there’s the same imperfect brickwall setup in our MSX then.

Didn’t FRS mention this earlier, that there was a second filter?

Yes, it’s very different, the YM3012 DAC receives the digital values for each channel from the YM2151 and then sums them before DAC conversion, so it needs enough bits to accomodate the summing, whereas the OPLL outputs them one-by-one as a sequence of pulses. So you absolutely can’t directly compare the nr. of bits.

When I brought this difference up earlier you said to me that as far as the reconstruction filter is concerned it amounts to the same thing :D.

If I had to compare bit-ness somehow; if all 11 channels of the OPLL would be summed by a DAC as well it would need 13 bits. The YM2151 needs more bits since it has to sum 8 channels with up to 4 carriers each, additionally I think its internal sine table is 10-bit. But I don’t think this difference in resolution is super significant.

Aah, nice :)
I'll have a look at all that.

Right. Tho these kinds of figures can mean anything. It could be a theoretical calculation that never actually happens in practice and it could be some sort of compound figure of which we don't know how it is composed.
So measuring is knowing as they say :)

Indeed, there are op amp low passes there as well.
But the values are a bit different from the SFG cart.
There is a filter right behind the RC filters of the opll which has a cutoff of about 6kHz.
And there is one after all the sound sources are mixed together which has a cutoff of 12.5kHz.
Since the cutoffs vary between the filters its not really a brickwall design. They just happen to use op amps for the filters.
Overall i would say that there is too much filtering going on from start to finish.

It was part of the other thread with the rurboR fix.
But i have mentioned it as well.
In fact, i have made a cutout of the mixer section from this document for discussion but i couldn't get it to upload. (where is the facepalm pengu!?) :)

Thanks for the insight about the mixing in the ym3012! Didn't know about the on-board mixing, tho now that i look at it it kindof makes sense considering the 9 bits bus between fm chip and dac. That would suggest the individual channels are 9 bits.

Well, once you bandlimit the signal after conversion (brickwall filter) they do become the same. In the end its about the information contained in the samples.
You can basically exchange time resolution (or frequency response, if you look at it from that point of view) for bit resolution (or dynamic range).
So basically wiggling a coarse bit fastly over time is equivalent to having an extra in-between value available but at half the time resolution.
This equivalence is expressed by integration, so filtering. And they only become the same with perfect filtering, which doesn't exist.

This is basically how DSD works and in general how sigma-delta conversion works.
These are basically single or multi bits (typically up to 4 bits per sample) but at a very high frequency.
DSD does a single bit at about 2 megahertz sample rate but is equivalent in information to 20 bits at 96kHz.
Basically, for every bit you remove you need to add a bandwidth of the same size as your wanted bandwidth.
So let's say you want this 20 bits at 96kHz then you could also do 19 bits at 192kHz, 18 bits at 288kHz, etc, all the way up to 1 bit at 1.9MHz. Hey! we arrived at the sampling rate of DSD :)

An upshot of this technique is that all quantisation noise goes up in frequency as well.
This means that for the integration of this super high frequency signal you can get away with a much 'shallower' reconstruction filter.

But all this is besides the point :)

In the case of the ym2413 we have to consider the 50kHz sample rate as well since it is a real samplerate.
Consider only one channel contributing a signal to the total with the rest being quiet. That means that there won't actually be a 447kHz signal anymore. You get a 50kHz (kindof pulsed) signal.
This then becomes the first sampling rate we encounter from out human hearing range perspective and so we need to deal with that frequency.
And once you cover this 50kHz case all the supersample rates of the channel slicing will also become integrated.

The significance in resolution becomes clear when you do the actual phase modulation. The limited resolution of the sine table will cause more temporal and amplitude distortion due to rounding errors, quantization noise and whatnot.
It can be (or rather, probably is) another source of noise which is separate from any noise the dac dynamic range adds.

Another significance (probably the main one) is that these higher bit depths lead directly to a lower noise floor. Generally speaking in audio you reduce the noise floor by about 6.02dB for every bit you add tot he sample word. For instance, CD has 16 bits, so the (maximum theoretical) noise floor will be 16 x 6.02 = 96.32dB from full scale.

My motivation for questioning the bitness of the ym2413 output was finding out whether this 9 bits we see could ever change. But it appears that this is a fixed thing (since the mixing is sequential and at a higher frequency), which is great because then a sharp reconstruction filter at the appropriate frequency (about 20kHz) would certainly be helpfull in some way.
Tho it may not be the only consideration given the possibility of internal noise due to the actual PM algorithm.
This reconstruction filter only deals with reconstructing the output. If the output signal contains intrinsic noise (so not the noise caused by the DA conversion) then you will still get lots of noise.
This is something to keep in mind.

This thread really interests me a lot. My Panamusement FM-PAC does not work anymore and I will try changing some caps sooner or later hoping to fix it!
I know, from YM2413 datasheet, that YM2413 has a 9 bit DA converter built in, that MO and RO outputs out of a source follower transistor, so output impedence is Rs/(gm*Rs + 1) with gm transconductance of transistor (unkonwn) and Rs source load, for simplicity sake, pretend Rs=R5=R6=2.2kOhm.
I know, from YM2413 application manual, that there's only one operational cell used twice per channel, there are 9 FM channels but served in time division single access logic arranged in 18 time slots, so with a clock of 3.579545 Mhz and needing to look into registers configuring FM channels, it comes out that generated sampling rate per channel is fM/72 that is 3.579545/72 = 50kHz (YM2413 application manual p.14).
It seems to me that Sega Master System/Sega Mark III YM2413 sound module has a design similar to Panamusement FM-PAC and other Panasonic MSX machines, especially looking at integrated circuit RC low pass filter for MO RO signals.
On the other hand, Daniel Tufvesson doesn't like the concept of "integrated circuit", as everybody else, and he get rid of capacitance, see The YM2413 FM OPLL and zoom at the proposed circuitry.

Excellent, thank you very much! I think it is a good improvement! Thanks for taking the time.

I’m looking forward to trying it out, glad I hadn’t ordered components yet (but will do soon).

Totally agreed on the constructive approach. I feel the same way.

I’ve updated it on the MAP.

Heh, he used my Xak VGM pack to test it, that’s nice :D. Dearly lacks an AY-3-8910 though! Amazing what difference that can make.

Actually, the DACs use a serial protocol :). But I’m not sure exactly how many bits the YM2151 uses, the internal workings of the OP* range operators are quite similar¹, so my first instinct says that between OPLL and OPM it’s the same, however I couldn’t explain the DAC needing 16 bits in that case. I looked briefly at some YM2151 emulation code and think I noticed 10 bits there, but could be wrong.

About the phase modulation, indeed the phase modulation can add “noise”, as you can see here:

However, that’s where the OPLL is truly no different from any of the other Yamaha FM chips. (They can also easily add aliasing I reckon.)

I’m not too convinced that noise is a big problem, at least not bigger than the noise problem of my MSX in general :D, but I will find out!

¹ Although OPN and OPL do form two fairly distinct branches from the original OPM; OPN is more similar to OPM, OPL originated as a low-cost version and then expanded in a quite different design. OPM/OPN have only sine waveforms but are all 4-op and with a choice of 8 algorithms, while the OPL series expanded by adding different waveforms, only introducing 4-op in the OPL3 and with only 4 algorithms (only 2 of which are truly 4-op FM).

sd_snatcher: the original drawing has an LPF block behind the RC filters which are directly on the output. (Sidenote: any idea why there are separate filters for both RO and MO?) So, I guess the 20kHz LPF that the document is speaking of is that LPF block. Why isn't that block in your revised drawing?

Is it because it doesn't make much sense to have that RC filter with low cut off frequency and then after that an LPF with 20kHz cut-off? Or what am I missing here? (There's so much text, I may just have skipped the explanation... sorry about that.)

There, “Yamaha YM2413 OPLL application manual with SFG-05 filter” uploaded .

And

The reason for this is that all music (from before emulator cross development) were made on chips with those exact filter settings/values. Whatever boost or weakening of certain frequencies they had were compensated for by the composers during the song creation process. Eg "oh crap this hihat/trumpet/string sounds too loud/obnoxious/hurting my ears.. lets lower the volume")

Ofcourse there were other factors like different devices/mixers/amps/speakers/room accoustics but I HIGHLY doubt that ANY settting/setup/accoustic would have made MSX of the composers sound like the fixed turbo R's sound, since you want your development environment or studio (if any) to reflect the user base, and nobody had a "fixed" turbo R back then.

In fact, I think if the OPLL sounded like "FRS-fixed" from the beginning, everyone from Microcabin to the dutch msx scene would have just made music for that one and adapted to whatever frequency imbalance it has. And nobody would have a problem with it because I'm sure tunes wouldn't have as much treble as the actual versions we have now.

So you can't compare a opll with changed filters to a YM2151. Yes they might sound the same now frequency-wise but close to 30 years of msx-music were made (again not counting those composed on emulator) for an opll without changed filters. So all those games would sound off imo.

First of all I primarily use BlueMSX for gaming due to easier/faster user interface, and I never had any problem regarding tone balance/frequency response or whatever its called, because it seems blueMSX seem to have less treble than openMSX. I only use openMSX for code testing and Trilotracker FM and while doing that the higher frequencies obviously wont be a problem since it is in my own hands sort of.

I just did a quick test (fray intro tune) on my PC, openMSX seem to have less treble than the FRS samples, while blueMSX has less than openMSX.
https://www.dropbox.com/s/tzhkvoy3w9yrdo8/fray%20test.wav?dl=0

Just a quick 'n dirty unedited recording of my desktop sound, with openmsx bluemsx and soundcloud open and playing them on and off. I messed up halfway but I don't feel like redoing it, since its not much work to test it yourself.

I'm only talking about high frequencies here, not about whatever detail instruments might have or not have, emulation accuracy or even whether the filter of bluemsx is correct or not.

I think openMSX is a bit too strong on the high end either, but I guess not many ppl notice it because not all games have as loud high leads as Fray (Like I said in the turbo r fix thread, maybe made on a turbo R ST with too much filtering).

I hope this explains my POV.

@syn

I don't mean this to be offensive in any way: The fact is, you know, that nobody has any obligation to fix their machines.

The idea is the exact opposite: those who feel that their TRs had terrible mixing and muffled sound, and those who had a variety of problems with their Sanyo2+s (including even muted sound cartridges) can finally have a solution to make them happy.

If you're happy with your machine as it is now, why bother?

@syn 1) Good, I feel so too! 2) I don’t think you got carried away… 3) Maybe the modified circuit¹ will suit your ears better, we’ll see. For the record, it’s easy to fix the turboR / Sanyo audio with original OPLL filter by picking 0.015µF caps (though I only found 0.010µF polyproplylene but that’s close enough). Your choice! 4) Well the balance will good .

¹ Now in the tR fix there’s three filters following the OPLL, all three with cut-off ~19K if I understand it correctly, so the slope should be -18 dB per octave now, pretty sharp.

So, let me respond in a general way.

I still feel the design document has been violated, but i guess at least now the original reference is included.
I think the appropriate way of doing this would have been to turn things around, so leave the document in its original form and add a document with the proposed changes separately.

About the proposed filter, i have checked it in a simulation and it does a very poor job of the reconstruction task.
High frequency waveforms look like triangle waves instead of sine waves.

I think that it is also a bad to assume that the filter in the SFG-05 is suitable for this case.
Differently specced chip with a differently specced DAC.

Then there is the whole business with the LPF block in the drawings. It was a total surprise to me because it was left out by sd_snatcher in a previous version. See where frivolous revisionism gets you? I now have 3 versions of this document...

I also played around with a filter design to make an 8 pole reconstruction filter. When applied at around 15kHz such a filter does indeed make a sine wave out of a 25kHz block wave. So it works pretty well as a reconstruction filter.

It requires 4 op amps (which come in single packages and cost about 2~5 dollar), 8 resistors and 8 capacitors.
So this circuit would cost about 10 dollars to make.

Unfortunately i don't have the required components at home (including sound chip) so i can't actually build it.
But it's dead easy to design such a filter and if you want a good quality enhancement of the sound then this is a much better way of doing things.

And what musician worth their salt would accept 3 squarewaves as their instrument? Huh?

This one requires a bit more background on how things went in this project.

All the motivation came in order to help people that have been complaining about audio problems in the Sanyo2+ machines either in the Brazilian community or later here on MRC.

I purchased one PHC-70FD just for that, and it arrived a bit before the thread on MRC started. Sadly, a few weeks later this machine started to malfunction and have all sort of glitches, up to a point that it didn't turn on anymore.

The problem seemed to be in the SMD components with tiny pins that are way beyond my soldering skills. So I paid FBlabs to fix it and reverse engineer its audio circuit. He did a splendid job in both cases. The T9769A had to be completely re-soldered because the old solder had cracks in it.

The reverse engineered schematics were a speed boost. And curiously, digging into bugs of either old circuits or old software has a kind of archeology aspects into it. Everything is interconnected, but you start to see layers of what happened into the project. In software, you clearly see things that were bugs in the library, because they're present in other software of the same company, and things that are bugs of the game code itself. Once you change something, you trigger another bugs what where latent in the code and have to deal with them. It's identical in the hardware world.

One day, just after I had finished the testing session of the Sanyo2+ fix, I went to use my usual FS-A1ST, and - my goodness - by hearing it just after the fixed PHC-70FD made it even more clear that it sounded like crap! I always knew it had a poor sound, but now it had become so evident that was impossible to ignore.

So I though: if I fix this problem too, I can also help a lot of people. I've seen people complaining about mixing and muffled sound problems all the time on MRC, and even trying to resort to software workarounds to get rid of it.

But the mixer calibration took an immense amount of time and testings to get everything fine tuned. Much more than I estimated in the beginning. After making any adjustment I had to come back and test the whole lot of music library again. This made me realize why the Panasonic engineers just skipped this step in the turboR design. They probably where in a project that was being rushed.

Once everything was finished, I though: it will be a waste if all this knowledge gets lost. And there are a lot of venturous new MSX hardware creators like ReproFactory, FBlabs, Tecnobytes, 8bitsforever etc that I can also help if I publish some kind of documentation about this.

And the existing YM2413 Application Manual was in terrible shape. There were Tables that had a lot of unrecognizable numbers, figures with a lot of problems that could lead to design errors. I've been slowly working for years in a cleaned up version, manly because it could be of great help for emulator developers. "I could join both efforts", I naively though!

Well, fixing an old document to me sounded as normal as fixing any of the old software or old hardware I already fixed. And over the years I have grown a thick skin for critics I sometimes hear like:

- "You should never have touched Metal Gear 2! You don't know if the designers intended it to be slow like that. A turbo has no place in this game."
- "You should have never touched Gradius-2! If the game programmers planned the game to have smooth scroll, they would have coded it like this from the beginning. And the colors? It's totally different from the pallet the artists have used to draw those graphics! You have totally ruined the gaming experience!"
- "Did you ever ask permission to ASCII to make any changes in the MSX-Audio BIOS? The inclusion of MSX-Music routines/instruments in it is wrong and unforgivable! This has unpredictable consequences that will end our civilization as we know!"

I obviously added a bit of fun in the expressions above to make this topic lighter, but it goes more or less like this in general lines. And some parts of it are literally true.

I never intended the fixed Application Manual to be passed as the original, and I never said that the original should be deleted. When I contacted Grauw to publish the new document on his site, I clearly explained what I did, and asked if he would be interested in adding the fixed version there. In no point in history I asked him to delete the old document or even crossed my mind that he would or should do this.

I even have put a large disclaimer just under the modified drawing. Anyone who wanted could look into the original document to see what was changed. I let it very clear who made the modifications in the document, so anyone who had any suggestions for improvement could easily contact me. I made a public announcement in a very known site to make it clear of what was being done, instead of trying to sneak it into some alldatasheet.com as original.

So I did the best I could to help my friends here in every possible layer of this episode. Be those who use the existing circuits, or those who work hard to create new circuits.

Ok, so maybe i misunderstood the intent. Tho the name of the document doesn't make it clear that it is a modified version.
All i'm asking is that, since the modification is based on personal opinion (maybe driven by personal situation, like the chain used to listen to the sound), it should be published separately, even as an addendum to the original document. As long as the original document serves its purpose as a reference to work against. This is the important bit. A reference, even when considered flawed, is still a reference.

Just to make it clear, i'm not against mods in general. I'm against making any mod a standard over the original standards since any new standard will be arbitrary (in the sense that there are multiple solutions which themselfs are based on personal preference) which can only lead to confusion of the unexpecting reader.

Also let me make it clear that i have no personal grudge against any of you. I'm fighting for the cause, not for getting into trouble with people

Hey!

First off, I understand your reasoning here, just like I also understand the sense of presenting the information in-context like FRS chose to do.

However I would like to mention that, the only way for things to go exactly as you want it is to do it yourself. Ultimately the one who takes the initiative and puts in the work gets the final say in what feedback is incorporated and how.

Interesting. It would be nice to see a bit more elaboration on that, with some illustrations perhaps, so I would have a little more to go on. Since if you don’t mind me saying, it seems nontrivial to to model the exact characteristics of this chip and the machine .

Btw, talking about no assumptions; since I’m in the process of ordering the components, with varying cap values to compare, if you have any particular recommendations about things that would be good to test (e.g. waveforms, frequencies), I could take them along. I have an oscilloscope.

It was literally resolved within hours. Just a slight, should not even be worth mentioning. I think you’re putting a bit too high of an expectation on people. Don’t expect that everything is flawless straight away. Don’t expect that everything is flawless period, actually. (No matter the source.)

Note it’s only early days, you’re still seeing a work in progress. There’s been no implementation feedback yet, while I think it’s best to judge from practice. As a result of that more changes could follow, it could take weeks, months, even years before a final conclusion is reached.

Firstly, it would not address syn’s feedback, which is about the naked timbre of the OPLL rather than the filtered one used in varying degrees on MSX (presuming it’s not due to some form of distortion). So from his perspective, this would still not be the right answer.

Secondly, your design could get critiqued about its complexity and cost, where FRS’s is easy to integrate, and perhaps better fits the “80s design”. Due to that modders and cartridge producers could favour FRS’s approach over yours.

Just saying this to point out that if people have different goals, with different trade-offs, there is perhaps not a single correct answer (though maybe a generally favourable). FRS takes his approach, syn’s goal is an easy modification of that, and there’s yours as well.

I’m serious though! At work we have an in-house audio engineer, although his budget is of course not infinite, he would definitely not settle for sub-par equipment. I have such high regard for the Micro Cabin musicians, I can’t imagine them being ok with that.

For the record the original was never deleted, it has always been there, just for a few days it was displayed less prominently. I considered this a better readable version with an error corrected (not uncommon for Yamaha). I participated in the discussion and updated things as we went to take new insight and feedback into account, just like I always do.

But I nor FRS should have any need to justify ourselves, so if you have any complaints about that then you can shove it up your… Pardon the language. Let’s say, maybe next time assume good faith.

By the way, note that although I’m intrigued by the potential of opening up the filter, since I’ve always found the sound of the FM-PAC lacking; since the MAP is about software development, the primary reason for hosting the document on the MAP is because of the clean-up. Were it purely the schematic, there are better places.

And it’s still there! Nothing is supplanted. FRS’s version and the significant changes relative to the original are prominently mentioned. I see it as no different than e.g. Eugeny’s version of the V9938 AM.

Only the users can make anything a (de facto) standard.

If you can tell me how to upload pictures then i could post some pics of my tests.

What i was aiming for was seing what the filter does with a 20kHz sqaure wave.
According to sampling theorem such a square wave should be reduced to a sine wave.

If you have a 20kHz square test signal you can do the test yourself.
BTW, i chose a 20kHz waveform because that is basically the highest point you would need the filter to do its work at.
Theoretically you could test closer to 25kHz but that has no real purpose. But you could test closer to 25kHz if you like.

About components, resistors and capacitors don't change a whole lot as long as you use proper type of caps and the values don't vary wildly. Make sure the caps are nonpolar, so in practice either poplypropelene or ceramic discs, depending on required capacitance. But it kindof depends on the actual schematic you want to implement.

If you also want to test the effect of the op amp then you should buy from a proper certified shop. Ebay and ali express will get you in trouble almost 100% of the time. So use a proper components shop like farnell, mouser, whatever.
For the filter proposed by sd_snatcher something like a TL071 should be pretty good. I think they are less than an euro in price and come in DIP packages.

Tho i'm actually unsure about the actual slew rate the op amp will require.
This is depending on the actual voltage swing the chip outputs.
It's going to be somewhere between 0V and 5V, but the question is how big the biggest possible swing is going to be within that range. From there one can calculate how much the op amp needs to be able to process.

In any case, for such a simple filter the TL07 range is going to be just fine. It's common, cheap and it has enough bandwith.

For my filters you'd need a better op amp because these filters are particular.
The tool im using calculates that i'd need about 20MHz of bandwith.
This is not very much in modern standards but it still is more than the 3MHz of the TL07 range, so its not going to work well with those.

Fortunately these op amps with the higher bandwidth are not that expensive either. But they are SMD, which may put some people off. They are available in SOIC packages, which are the biggest of the SMD packages so it's not all that bad.
And you can get them in quad, so 4 opamps on one IC.

It's even worse then that.
For the same goals there are already several possible solutions with maybe minimal differences.

Like i said, i used a tool to design a (actually several) filter.
It is really really trivial to use this tool. I could poop out a filter with different specs every 2 minutes. Designing the filter is actually pretty easy if you have certain specs in mind. You actually get to see a simulated response curve and you get to choose from a range of op amps. You even get information about how much it will devfiate if you use particular precisions in resistors and capacitors.
I could choose the cutoff frequency, stopband frequency and stopband attenuation. I could chose different optimizations and different topologies.
And all these choices just for a single 8 pole filter. So it's kindof difficult to settle on a particular design, just because of the amount of choice and the minimal influence on results :)

But these are just the basic filters which are not yet tuned to interface with the ym.
So i am actually a bit uncertain about giving you guys a schematic. I'm not sure myself what the best solution would be because i don't have the chip.
I could make some good guesses but someone would have to get the components and verify it.
Tho i'm pretty sure that they would do their work well as a reconstruction filter.
I have simulated some of these things and they seem to be working beautifully, giving nice looking a sine wave on the virtual scope (when applies to the mentioned 20kHz square wave).

BTW, i'm planning on buying some op amps soon, some of which are suitable for such a filter. If i get them i may do some tests to show how the filter operates on real world test signals.

Yeah, but this sounds like an argument for leaving the filtering alone!
In fact, i almost used it myself.
See, if extending complexity means that you will get too much high frequency content then the obvious thing to do is to filter some of that high freq energy to to make the sound overall more balanced.

You just gave a very nice motivation to have the top end be reduced in some way.
Not sure if that was originally thought up by the yamaha engineers, but it sure does work out that way.
It's part of what syn was arguing before.

In fact, it does matter a lot. As I explained before, 80's solutions were smart solutions designed to take the best possible outcome from the least possible cost. The video/sound chips themselves were far from perfect, so nearly all solutions relied in some tricks and shortcuts to make them seem very good, while still being able to be affordable.

There weren't millions of simultaneous colors at that time. There weren't huge power supplies that can provide almost limitless current. There weren't gigabytes of RAM. PCM chips still sounded much like crap for today's standards. You'll see severe constraints on resolution, DAC and analog circuitry everywhere you check on an 80's computer.

That said, it's a know fact that a perfect human hearing goes up to around 20 kHz. But it's not all that known that above 10 kHz the resolution of the human ear becomes precarious. Yes, we can hear the tones. But normal humans are incapable of discerning an sine wave from a triangle wave above 10 kHz.

Whoever doubts this statement can try this test for themselves here. Do a blind test with your friends to see what's the highest pitch where they can still tell you what is the waveform you're selecting. Ask someone to do the same blind test on you.

As the analogy I used before, worrying about artifacts above 10 kHz is like worrying about a scratch made by a grain of sand in the lower side of the vehicle. Sure, the scratch might be there. But will any owner will spend a penny to solve something that no real humans can notice? But what anyone will clearly notice, obviously, is if the volume of the frequencies above 10 kHz are attenuated by something like an LPF filter with a cut-off frequency of 2.2 kHz. That's a huge bump in a area that's very noticeable.

Bonus fact: only kids an teenagers can hear up to 20 kHz. Gifted adults will go up to 18 kHz, and the majority of the adult population just can't hear over 16 kHz.

Now here comes the point I was trying to explain since the beginning. Step noises live in the domain of amplitude, not in the frequency domain. This means its a quantization error, not a sampling error. If you're trying to apply the Nyquist–Shannon sampling theorem to filter out quantization errors, I'm sorry to tell you that you're using a excellent tool for the wrong job. It's like trying to fasten screws with a plier.

Artificially lowering the brick wall filter cut-off frequency to smooth quantization noise just to please personal tastes, cannot be called "Nyquist–Shannon names sampling theorem" anymore and can be considered just as wrong as a cut-off filter that has a too high cut-off frequency. So anyone proposing this has no moral ground to state that others are wrong.

Yes, I know the theory very well. And I'm telling you since the beginning that a Nyquist–Shannon reconstruction filtering is not the tool that will smooth what you have been calling "haze and noise". You're shooting at the wrong target. And what you have been calling "haze and noise" is much louder than the barely human-noticeable distortions above 10 kHz you're wanting to spend big resources to clean.

But, again, you can prove that you're right and that you can make the OPLL output perfectly smooth and free of "haze, aliasing and noise" with just a filter that strictly follows the Nyquist–Shannon theorem as you said. You just have to build it. Then play the soundtracks of Starship Rendezvous, Valis-2 and Fighters Ragnarok in it to prove your point.

I'm not asking this in an offensive way, but have you actually ever dealt with OPL FM synthesis or emulation yourself? Have you tried to create any instruments?

I can tell you that if the composers ever wanted smoother tones, they had plenty of more conservative feedback values to use. It's just like the distortion pedal of an electric guitar. Someone just don't need to crank the distortion level all the way up, then apply some strong filter in the output to try to get a clean acoustic guitar sound again. It just doesn't make any sense, since they can just lower the distortion level from the beginning. If you don't like distorted electric guitars, you just have to select music from a genre that suits your taste better. Or just patch the damn games to replace the instruments you don't like.

Maybe it was my fault to select songs for the demonstration that made heavy use of the feedback register, and even some that use the PSG square waves to further increase that. I should maybe have selected more songs from Compile and Bit2, as those are very conservative and mostly only produce way behaved clean sounding instruments with the OPLL.

Another thing that you don't seem to be aware is that every experienced OPL composer knows that when you set the feedback register up to its maximum, what you have in the channel output is white noise. This is done intentionally to produce drumkits just like the white noise generator of the PSG is used for that purpose.

And you know what happens when you apply a low-pass filter with a low cut-off frequency to white noise that the composer intended to be sound like their cymbals or hi-hats, right? You just destroy the white noise, and the cymbals/hi-hats will sound like they were being player from a 30 years old cassette tape that was left over the loudspeakers this whole time.

Enough of throwing pearls to pigs.

Going down to your level of argumentation, do whatever you want with your ANALog filtering. I'm sure you can extract perfectly round waves from the OPLL as much as it's possible to use analog filtering to extract perfectly smooth RGB ramps and also smooth line-art from a V9938⸮ They'll look great on the oscilloscope and I'm pretty sure there won't be any lack of detail as a side-effect⸮

I'm again sure that you also have been seen before in these forums using the electronics knowledge and means that you have been proudly demonstrating to possess, together with the same level of energy you're using in this thread, to help those who post threads about problems in their machines . That speaks a lot about your personality.

My conclusion about this whole matter is pretty clear. I don't have time this crap anymore. My priority right now is my rehab, and its huge amount of exercises. They require a great deal of concentration, so I cannot afford to worry if some socially challenged fundamentalist is online stomping on my reputation and making ridiculous of my decisions. It's impossible to keep your focus when you're worrying about something like this.

Given this new situation, I've been growing increasingly worried about releasing the projects I finished before my surgery. They modify existing hardware/software, so there's always space for harsh and mindless criticism. One of the projects that are worrying me the most is the documentation about the de-facto standard for HID devices on the MSX, that ASCII and later Sega of Japan followed, and that joyTest and newTRdrv takes advantage.

While the standard works like a charm and allows easy detection of the connected devices with minimal resources, it's still just a de-facto standard and it's not perfect. After someone takes the risk of turning tacit knowledge into explicit knowledge, there's always space for some fundamentalist just come out of nowhere to make ridiculous of the whole thing for whatever reason.

I thought I could help others here to take advantage of this standard by documenting it, but the risks right now seem clearly too high for the rewards.

I still have to give some thought on what I'll do with the rest of the projects. While my emotional side feels that it's sad to have invested all this time to create something that won't ever see the light, the rational me thinks that it would be simpler/quicker to just to bury everything under some zip in my HDD and live my life free of those concerns.

sd_snatcher, please keep up the good work! I'm sure we're all very excited about your projects. Even redman, syn, me and other 'fundamentalists'....

It's just.... If I may use some french words: 'C'est le ton qui fait la musique'. In the newsposting, it was blundly stated that there is an error in the application manual, that it has been corrected and that the new manual can be found there and there. And at least in the beginning of the discussion, it also looked like that the original document was replaced and there were no signs in the new document that there had been some alterations.

Later on, things proved to be different, but that WAS the initial tone that was used. And from there, the feedback and the discussions started.

If you would have brought the news as: 'Hey guys, I found out the value of the capacitors of its twin low-pass filters can be changed to 20khz to have much richer sound. I decided to make an appendix to the original manual, which I also refurbished because of readibility issues', you would have been crowned as the new king and this thread wouldn't have had more than 160 reactions....

Anyway, foremost, I'm really really glad your operation went well and I deeply hope you will recover completely!

Understood, Grauw, but blundly stating there is an error in the Yamaha document is just as aggressive as the sentence 'how do you have the nerve to touch it'. Action provokes reaction.

Please guys, can we just all have good make up sex and continue our MSX love ^__^

Well, I had no problems with the tone of Snatcher's post. Hey, we all make mistakes or bugs, but that was not intentional and Snatcher did not use any hard words on the manual authors (who will probably never read this thread anyway!).

I was way more disturbed by the tone of the reactions here about a work researched and published for free and for the community. This might just lead people that are doing and creating things here (and I insist on the "doing and creating", wether it's hardware or code), to keep it for themselves or just a few friends, in order not to be exposed to such reactions...

Sorry to hear.

.

Look, there is nothing wrong with releasing mods and whatnot.
Mods are fun.
As long as they are separate from the original documentation.

I don't want to discuss this further due to your condition. Seriously, it's all good.

No, i'm not kidding. I see no button/option to upload pictures. Or audio files.

Been looking for one for hours yesterday.
Originals are around $200,- it seems.
Can you still buy FM Stereo PAK somewhere?

Besides that i don't have an msx with a disk drive anymore.
I'll think of something.

Can't be done directly. Some peep use https://www.msx.pics/ for pictures or use something like https://imgur.com/

Oh! I did a math error, entering data values, so, as it has to be, cut-off frequency of low pass filter is 20 kHz.

Thanks, i suspected something like this but its good to know for sure that i didn't miss any button.
Typical that my usual picture site (photobucket) recently decided that they want money for hosting/linking pics.

I think at least in case of FRS’s fixes, since the goal is to improve sound from existing machines and cartridges, a constraint is to be able to easily retrofit it on existing board designs.

Of course depending on your goals, there’s no need to feel tied to this constraint.

As for "spirit", I think it’s mostly about tricks, hacks and trade-offs they would’ve typically made back in the 80s to keep circuit complexity and component cost down to levels which they felt were reasonable back then. Like, if you can e.g. rely on limited op-amp response for high frequencies, saving you components further down the line, or if unwanted frequencies can be reduced to levels acceptable for users or hardware cheaply instead of fully eliminated, then it would’ve been something they would’ve perhaps done.

Yep, some machines have stereo output, e.g. the Palcom.

Gentlemen,

Since we all are looking for the YM2413 reference, I have assumed that Yamaha themselves knew how to use that chip, i.e. the music products were designed and tested to the spec by Yamaha. I don't expect any errors in the schematics of the Yamaha products using YM2413.
Now, we take two products:
- PSS-270
- PSR-6
Both use YM2413 and both have the same active filter with almost the same components.
Here I have captured the schematics of the output filter:
Equivalent schematics
I use LT1677 opamp in place of JRC4558 for the sake of simplicity. Also I have used the source follower as the output stage of YM2413, since originally it was the NMOS device.

And then, I've performed the AC analysis of this circuit:
AC pSPICE analysis

And finally as the result of pSPICE simulation, here are the graphs of the frequency response for the drums and music output:
Frequency Response Curves

As you could see, the gains and cut-off frequencies are different for the drum and music channels. Also the transfer function for drums rolls off starting from 1 KHz, whereas music channel starting it's rolling off from 3 KHz.

I also noticed, that despite the fact that MO and RO have the same filter stages the responses differ. IMHO, this is due to the differentiator (DC block) stage being placed in between in a different fashion (came to the same conclusion).
I can try simulating with the OPAMP with lower input impedance, however this shouldn't be the problem: JRC4558 has 5MOhm of input impedance.

Also the rhythm section acts as a TWO POLE LPF. The roll off of rhythm section is clearly 12dB per octave/40dB per decade as we look at the original schematics.
The music output is also TWO POLE LPF, but the cut-off frequencies are different (don't overlap). Is this deliberate? PSR-6 and PSS-270 are slightly different in passive components:
PSR-6 (PSS-270):
C2=10nF (4.7nF); R2=15K(12K); R5=68K(82K); C5=10nF(4.7nF)
These two poles are formed for music section by the RC filter R3/C2 and the second active stage R5/C3.
For the rhythm section -> R7/C5 and the same R5/C3.
As we can see, the gain of the music section is higher for PSS-270 : gain=-6.8 vs PSR-6: gain=-4.5
and for the rhythm section, the gain is resprectfully for PSS-270: gain=7.8 and PSR-6: gain=5.5.
PSS-270 has 3.4KHz cut-off frequency for the first stage and PSR-6 has 1.6 KHz.
Also, the cut-off frequency for the opamp stage is 8.8KHz for PSS-270 and 10.6KHz for PSR-6 .
However, I don't really understand why Yamaha guys decided to cut-off the low end of the music section making a band-pass filter. Maybe removing the low frequency flicker noise? At the same time, the way how the circuit is biased is not optimal from the noise point of view (resistors R4 and R11).

I have assumed that the output driving stage is a source follower. Then, the output impedance will be defined by the source resistor of 2.2K. Maybe I'm fundamentally wrong and the output stage is push-pull. In this case, on the modeling schematics we can depricate the R1 and R6.
I'll try to re-run the simulation with 0 Ohm R1 & R6 resistors (and the high pass will be removed).
Also I can run the frequency sweep on transient (not AC) analysis and get the max amplitude and plot response this way.
I'll publish the results tomorrow.

@redman: your suspicion were correct, the circuit wasn't biased properly for the rhythm section (and autobiasing was off). See corrected response here.
And here is the updated schematics for the AC analysis.
Now the rhythm section and the music section agree very well.
@Giangicomo Zaffini 2:
I agree with you. But since the output NMOS transconductance is unknown (we have to see the decapped die to find the W/L ratio of the output transistor), I have assumed the worst case of Ro=Rs=2.2K. In reality the output impedance is much smaller. I've simulated with different output impedance values, so the roll-off frequency slope is offset by a few dBs -> variance is irrelevant.

I compared the filters from the FS-A1WX with the ST, GT and FRS’s audio fix:

1. First OPLL filter:

WX: 4.7 kΩ / 15000 pF = 2258 Hz
ST: 4.7 kΩ / 33000 pF = 1026 Hz
GT: 4.7 kΩ / 33000 pF = 1026 Hz
FRS: 4.7 kΩ / 1800 pF = 18813 Hz

2. Second op-amp OPLL filter:

WX: absent
ST: 56k / 470p = 6047 Hz
GT: 56k / 470p = 6047 Hz
FRS: 56k / 150p = 18947 Hz

3. Third general audio filter (after final mix):

WX: 47 kΩ / 270 pF = 12542 Hz
ST: 47 kΩ / 270 pF = 12542 Hz
GT: 39 kΩ / 560 pF = 7287 Hz
FRS: 47 kΩ / 180 pF = 18813 Hz

So clearly the GT has much more filtering than the WX or even the ST (quite an extreme difference), and the WX gives some minimal values for the filters which are much higher than the GT already. For those who want to make a conservative GT audio fix, they can base it on these cutoff frequencies.

FRS’s fix cranks these up a bit more to the levels used by the SFG-05; three octaves for the first stage, introducing a second filter stage compared to the WX for a stronger fall-off, and half an octave for the third stage. So there is a -18 dB / octave filter at 18.8 kHz rather than a -3 dB / octave filter at 2.3 kHz and another -3 dB / octave at 12.5 kHz.

So overall relative to the GT it is a big change, but relative to the WX actually not as much… As you can see above, to make FRS’s fix match the WX values, change the first two caps to 15000 pF, remove the third cap, and change the fourth cap to 270 pF. I’ll be using these as a baseline for my tests and then compare that with FRS’s values. I’m looking forward to hearing the SFG filter on the OPLL!

p.s. I just submitted my order for the components .

I posted a review with comparison recordings.

Double posting Gyabuneko tweets on fixing Panasonic turboR MSX-MUSIC path here.